Gait training is pivotal in accessing the degree of recovery in walking ability of patients after treatment. Potential gait issues must be detected early in order to prevent falls while walking. In order to enable convenient and continuous gait training, we propose Gaitroid using multimodal sensors and smartphone to enable self-guided training via adaptive audio feedback and sensor-based gait assessment. Besides assisting user in exercising, gait phases and features useful for assessment. Initial experiments on evaluating gait phases detection and abnormality recognition with healthy subjects achieved average accuracy of about 90%. With Gaitroid, users can continue gait training conveniently at home while allowing therapists clinically evaluate gait performance and manage personalized therapy remotely, thus shortening the recovery progress.